I am stuck in a peculiar problem where I need to know the correlation between attributes. So far it has been easier, where all of my input attributes used to be only numeric, and hence without any hesitance, I used to generate a Pearson coefficient based correlation matrix. But now I am working with a business stakeholder, and they have provided me a data that has:
2 Discrete numeric attributes
2 Ordinal attributes (non-dichotomous)
5 Nominal attributes (dichotomous)
Following are the datatypes:
Attribute 1: Numeric (skewed)
Attribute 2: Numeric (skewed)
Attribute 3: Ordinal
Attribute 4: Ordinal
Attribute 5: Nominal (dichotomous)
Attribute 6: Nominal (dichotomous)
Attribute 7: Nominal (dichotomous)
Attribute 8: Nominal (dichotomous)
Attribute 9: Nominal (dichotomous)
I googled and found that for different kind of pairs, I need to have a different method and thus a different scale each time, to judge whether a pair of attributes are correlated or not.
Numeric and Numeric: Pearson method
Numeric (skewed) and Numeric (skewed or normal): Spearman method
Ordinal and Numeric (skewed or normal): Spearman method
Ordinal (non-dichotomous) and Ordinal (non-dichotomous): ?? (I am not able to figure out this one)
Ordinal and Nominal (dichotomous): Chi-square based Cramer's V method
Nominal (dichotomous) and Nominal (dichotomous): Chi-square based Cramer's V method
Nominal (dichotomous) and Numeric: Logistic/ANOVA/Point-Biserial
At the end of the day, I wanted to produce something like this, where I can easily explain which attributes (categorical or numerical) is correlated with others, in just one place.
Any correction, best-practice, comment or suggestion is most welcome.
